How to efficiently and accurately predict the channel impulse response (CIR) is crucial to wireless communications. In this talk, we first briefly discuss the prevailing learning-based approaches and the key problems involved. Then we introduce our work on the CIR representation and prediction with only the user's position information and a set of channel instances obtained within a certain wireless communication environment. Specifically, we resort to a novel physics-inspired generative approach to design the learning network, which makes use of the physical model of EM-wave reflections along each path and the excellent properties of the Gaussian Radial Basis Function network (GRBF) and sinusoidal representation network (SIREN) in predicting the path amplitude and phase respectively. We also discuss how to extend to the resultant learning architecture to the MIMO case and how to apply it to mobile channel prediction.